Remove all traces of low_cpu_mem_usage (#38792)

* remove it from all py files

* remove it from the doc

* remove it from examples

* style

* remove traces of _fast_init

* Update test_peft_integration.py

* CIs

docs/source/en/internal/model_debugging_utils.md

@@ -51,7 +51,7 @@ torch.random.manual_seed(673)
 # load pretrained model and processor
 model_id = "llava-hf/llava-1.5-7b-hf"
 processor = LlavaProcessor.from_pretrained(model_id)
-model = LlavaForConditionalGeneration.from_pretrained(model_id, low_cpu_mem_usage=True)
+model = LlavaForConditionalGeneration.from_pretrained(model_id)
 
 # create random image input
 random_image = Image.fromarray(torch.randint(0, 256, (224, 224, 3), dtype=torch.uint8).numpy())

docs/source/en/llm_tutorial_optimization.md

@@ -236,7 +236,7 @@ flush()
 Let's see what peak GPU memory consumption 4-bit quantization gives. Quantizing the model to 4-bit can be done with the same API as before - this time by passing `load_in_4bit=True` instead of `load_in_8bit=True`.
 
 ```python
-model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, low_cpu_mem_usage=True, pad_token_id=0)
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, pad_token_id=0)
 
 pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
 

docs/source/en/model_doc/chameleon.md

@@ -170,7 +170,6 @@ model_id = "facebook/chameleon-7b"
 model = ChameleonForConditionalGeneration.from_pretrained(
     model_id,
     torch_dtype=torch.bfloat16,
-    low_cpu_mem_usage=True,
     attn_implementation="flash_attention_2"
 ).to(0)
 ```

docs/source/en/model_doc/llava_next.md

@@ -157,7 +157,7 @@ import requests
 
 processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
 
-model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16, low_cpu_mem_usage=True)
+model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16)
 model.to("cuda:0")
 
 # prepare image and text prompt, using the appropriate prompt template
@@ -292,7 +292,6 @@ from transformers import AutoModelForImageTextToText
 model = AutoModelForImageTextToText.from_pretrained(
     model_id,
     torch_dtype=torch.float16,
-    low_cpu_mem_usage=True,
     use_flash_attention_2=True
 ).to(0)
 ```

docs/source/en/model_doc/llava_onevision.md

@@ -121,7 +121,6 @@ processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-
 model = LlavaOnevisionForConditionalGeneration.from_pretrained(
     "llava-hf/llava-onevision-qwen2-7b-ov-hf",
     torch_dtype=torch.float16,
-    low_cpu_mem_usage=True,
     device_map="cuda:0"
 )
 
@@ -286,7 +285,6 @@ from transformers import LlavaOnevisionForConditionalGeneration
 model = LlavaOnevisionForConditionalGeneration.from_pretrained(
     model_id,
     torch_dtype=torch.float16,
-    low_cpu_mem_usage=True,
     use_flash_attention_2=True
 ).to(0)
 ```

docs/source/en/models.md

@@ -148,11 +148,6 @@ You need enough memory to hold two copies of the model weights (random and pretr
 
 Transformers reduces some of these memory-related challenges with fast initialization, sharded checkpoints, Accelerate's [Big Model Inference](https://hf.co/docs/accelerate/usage_guides/big_modeling) feature, and supporting lower bit data types.
 
-### Fast initialization
-
-A PyTorch model is instantiated with random weights, or "empty" tensors, that take up space in memory without filling it.
-
-Transformers boosts loading speed by skipping random weight initialization with the [_fast_init](https://github.com/huggingface/transformers/blob/c9f6e5e35156e068b227dd9b15521767f6afd4d2/src/transformers/modeling_utils.py#L2710) parameter if the pretrained weights are correctly initialized. This parameter is set to `True` by default.
 
 ### Sharded checkpoints
 
@@ -245,7 +240,7 @@ Big Model Inference's second feature relates to how weights are loaded and dispa
 
 Both features combined reduces memory usage and loading times for big pretrained models.
 
-Set [device_map](https://github.com/huggingface/transformers/blob/026a173a64372e9602a16523b8fae9de4b0ff428/src/transformers/modeling_utils.py#L3061) to `"auto"` to enable Big Model Inference. This also sets the [low_cpu_mem_usage](https://github.com/huggingface/transformers/blob/026a173a64372e9602a16523b8fae9de4b0ff428/src/transformers/modeling_utils.py#L3028) parameter to `True`, such that not more than 1x the model size is used in CPU memory.
+Set [device_map](https://github.com/huggingface/transformers/blob/026a173a64372e9602a16523b8fae9de4b0ff428/src/transformers/modeling_utils.py#L3061) to `"auto"` to enable Big Model Inference.
 
 ```py
 from transformers import AutoModelForCausalLM